Case Study

The MKV Project: Summary Of Work

How we built an offline repair tool for an electric tractor between February and August 2026: the goal, the problems we met, and how AI and our years of software development and design work got us past them.

Covers February to August 28, 2026. The counts below cover the version-controlled work from May 24 to August 28.

Background

A tractor whose maker closed

White Rock Farm is a sister company of White Rock Future Tech LLC; the two share the White Rock name. The farm bought a Monarch MK-V, a battery-electric tractor made by Monarch Tractor of Livermore, California. It was a large investment in a new kind of machine, and it depended on the maker for support: dealers, parts, and the service software that reads the tractor’s fault codes.

Around the end of 2025 Monarch Tractor stopped responding to its customers, and in April 2026 it shut down. It laid off its staff and sold its technology, but not its tractor business, so owners were left with no factory support and no one maintaining the service software. The tractor itself still worked and had years of useful life left in it.

The farm could retire a working machine, or find a way to keep it running. It chose to preserve the investment, and asked us to build the means: a way for a technician to read the tractor, understand its faults, repair it and bill the work, without depending on a company that no longer exists. That is the project this paper describes.

The goal

One offline product, from tractor to invoice

Build one working product that a technician can carry to a Monarch MK-V electric tractor on a Windows 11 laptop. The product reads the tractor’s CAN bus, decodes what it finds, records the repair (findings, parts and labor), and hands that record to the head office, which turns it into an invoice. It has three parts: the repair application and its database, the field interface, and the CAN capture tool. They must work as one thing from start to finish, fully offline once installed.

Two design decisions shape everything else. First, the application is a repair platform that is not tied to one make of equipment, and the MK-V is the first library that plugs into it; a different tractor, baler or mower would be a second library, not a rewrite. That makes it an independent multi-brand repair tool, the same category as any shop’s scan tool. Second, the high-voltage side of the tractor is permanently out of scope, and a live working tractor is only ever read, never written to, except under a written safety procedure on a bench machine.

At a glance

The scale of the work

MeasureCount
Commits, repair application and database973
Commits, CAN capture tool151
Commits, field interface (vCab)171
Commits, engineering standards and records165
Database migrations, each reviewed and approved268
Automated tests, CAN capture tool alone1,981
Fault codes reconciled into one catalogue373
Entries in our AI defect and cure register362

The issues we faced

Seven problems and how we overcame them

The data had no history

The starting database held real tractor data, but the steps that loaded it were never written down and the loader scripts were lost, so no one could prove where a row came from. We set a standing goal to make the whole database reproducible: structure, stored procedures, data, keys and defaults. We rebuilt the history from migration ledgers, backups and version control, and recorded what could not be recovered rather than inventing it.

Three sources of fault codes disagreed

We reconciled them into one catalogue of 373 codes, ranked which source wins when meanings conflict, and showed that one file assumed to be an original was derived from another. Meaning is now read from one resolved view, never from the raw tables.

The field laptop could harm a live tractor

A capture tool that joins a bus the wrong way can put error frames onto a running machine. We wrote a seven-gate write-safety procedure (firmware first, repair before write, operator acknowledgment, bounded values, restore rather than alter, full audit) and built the capture tool to listen silently and refuse a bus it cannot make silent.

Real hardware behaved differently from the bench

A field laptop showed six CAN channels and the tool refused all of them on the strength of the first. We fixed the channel check, then found the fix had not reached the user because an older copy of the program was still installed in a second location. Every build now prints its own version stamp in its log, so the version in the field is a fact, not a deduction. On August 5 the compiled Windows program replayed a real 16,384-frame capture, and the output matched the input frame for frame.

The pieces were wired together the wrong way

The field interface had been given direct database access. We ruled that the field side talks to the repair application only by messages and holds no database credential, which removed a shared password and 23 hidden database codes from the field software. Fourteen design decisions on that messaging were put to the owners and ruled.

Money and records had to be right

A schema review found 69 links that recorded who made a change but were set to erase that record if the person was deleted; 73 links now refuse the delete instead. Customer credit became a payment method, and we fixed two defects: voiding did not return a credit, and overpaying created a new one.

Passing tests did not always mean working code

This was the most important lesson. Several times a full test suite passed while the thing it guarded was broken: test data that never reached the branch under test, a check that scanned an empty list, a reset tool that had not run since a schema change. We changed how we test. Every new test is first run against the unfixed code and must fail before it is trusted, and safety checks are mutation-tested: we break the code on purpose and confirm the suite catches it. After that change, 13 deliberate breaks to the bus-safety check, which had all passed unnoticed before, now fail the suite.

How AI and experience divided the work

What the AI did, and what experience did

AI did the volume: reading thousands of pages and files, writing code and database migrations, building tests, and working out undocumented file formats, all on our own server, with no source code or tractor data sent to an online service.

Our years of software development and design set the direction and caught what the AI missed. We decided the platform architecture, the safety rules and the data model (normalized, soft-deleted, under version control), and the rule that only the database owner touches the database. We insisted on visual sign-off, because a passing test is not a working screen. And when the AI reported something as done that was not, we required each mistake to be recorded with a lasting cure. That register now holds 362 entries, and the cures that matter most are built into code and checked automatically, not left as good intentions.

Who We Are

White Rock Future Tech LLC provides AI-related services for small to large businesses. We also offer CIO/CTO services on a part-time and consulting basis. Our company uses AI in our daily work, which is kept in house on private networks when dealing with client materials and IP. We check each claim from our AI agents against their citations as well as checking for supporting information.

AI is a potentially game-changing tool for business today. We believe that the use of AI, in the correct hands and with appropriate guardrails, boundaries and rules will be as transformative as the introduction of the Personal Computer (PC) was back in the late ’70s and early ’80s.

Find out what your project really takes.

Book a free 30-minute consultation. You will leave with a realistic sense of size, cost and risk, whether or not we work together.